A regularized data fusion approach to multisensor surface reconstruction and edge detection applications

Enhancing a robotic system's inspection and manipulation capabilities requires the extraction of numerous features from an environment. To retrieve multiple properties necessitates multiple sensors since each sensing device is limited to the amount and type of data it can incorporate.

In this thesis, we examine the problem of extracting given features, such as object boundaries, from different sensing modalities for recognizing objects. The mathematical analysis will focus on the early vision problems of surface recon- struction and edge detection. Since these applications are characterized mathe- matically by ill-posed properties, regularization is proposed as the basic tool for restoring stability and well-posedness to these problems.

This thesis extends Salinas' [29] prior doctoral work on regularized fusion. A fusion functional for integrating data from different knowledge sources will be formulated to overcome the difficulties associated with using only one stabilizer. This new hybrid functional, containing first- and second-order stabilizers, will give rise to filtered, interpolated surfaces and good detection/localization properties for the edge detection scheme. The algorithm extends standard regularization theory by allowing multiple inputs of different sensory cues, constant and variable smoothing parameters, and a weighting factor associated with each data source. The computational aspects of the algorithm as well as its performance on real and synthetic images are evaluated.